ValueError: Unsupported Ops in the model before optimization Dilation2D

[jdariasl]

Hi

I'm trying to convert a frozen model using tensorflowjs_converter but I got this error:

Traceback (most recent call last):
File "/home/julian/.local/bin/tensorflowjs_converter", line 8, in
sys.exit(pip_main())
File "/home/julian/.local/lib/python3.7/site-packages/tensorflowjs/converters/converter.py", line 671, in pip_main
main([' '.join(sys.argv[1:])])
File "/home/julian/.local/lib/python3.7/site-packages/tensorflowjs/converters/converter.py", line 675, in main
convert(argv[0].split(' '))
File "/home/julian/.local/lib/python3.7/site-packages/tensorflowjs/converters/converter.py", line 658, in convert
weight_shard_size_bytes=weight_shard_size_bytes)
File "/home/julian/.local/lib/python3.7/site-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 395, in convert_tf_frozen_model
weight_shard_size_bytes=weight_shard_size_bytes)
File "/home/julian/.local/lib/python3.7/site-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 144, in optimize_graph
', '.join(unsupported))
ValueError: Unsupported Ops in the model before optimization
Dilation2D

My environment:

tensorflowjs 1.7.2
tensorflow 2.1.0

[Whiax]

Same problem here:

Uncaught (in promise) TypeError: Unknown op 'Dilation2D'. File an issue at https://github.com/tensorflow/tfjs/issues so we can add it, or register a custom execution with tf.registerOp()

Model trained on Pytorch, converted to ONNX then to TF then TFJS/JSON.
It works on ONNXJS for CPU, but doesn't on ONNXJS with WEBGL backend. I wanted to give TFJS a try but I'm stuck at this error. If it can't be fixed, how should I retrain my model ?
Note that tensorflowjs_converter did warn me that Dilation2D was unsupported but I used the skip_op_check flag.
If you need a model using this op, I can share it.

edit: maybe it's linked with the maxpooling layer?.

tfjs 1.7.2
tf 2.1.0

[ai-bites]

I am facing the same problem. Could you please suggest a work around till a fix is released. Thank you.